


Cofounder, @LeashBio.


“Wet-lab validated.” AI binder design papers keep reporting impressive hit rates. But does it bind in a second lab? One independent retest: only 29/107 published AI-designed nanobodies showed binding—and 6 of those had high aggregation. Germinal stands out for the wrong reason: just 1 of 12 published binders showed detectable binding on retest—a result that warrants an explanation. How much trust should we put in headline hit rates? Read more on my blog 👇 smiles724.github.io/blog/ai-…

Claude was trained on protein sequences at scale Blog: xinmingtu.cn/blog/2026/claud… Code: github.com/XinmingTu/LLM-bio… Site: xinmingtu.cn/

We are delighted to share our new paper. By combining single-cell transcriptomics, live embryo imaging and functional experiments, we built a stage-resolved roadmap of mammalian implantation. And the greatest surprise? See tweet 5! nature.com/articles/s41556-0…

Self-copying entities filling unexplored ecological niches strikes me as *incredibly* likely
I think it's pretty likely that in 6 months there will be rogue AIs running around on the internet, appropriating whatever compute they can hack into or pay for (with stolen money or by renting out their own labor to assist with illegal activities) in order to run more copies. I also think this will be mildly annoying in the same way that the common cold is mildly annoying, but basically fine. Your smart TV will be weirdly slow sometimes, and a lot of crypto will get stolen, but people will get used to it. These AI-powered viruses be selected in the same way biological viruses are for not hurting their hosts too much and just making more copies. They won't have any incentive to damage critical infrastructure like electric grids. Then again, they also can't be held accountable for crimes, and will damage lots of old, poorly maintained, and insecure systems. It might be extremely hard to root them out once they've spread far enough, making the internet "radioactive" to connect to without hardened defenses. (To be clear I'm not talking about AIs powered by inference on centralized datacenters that execute actions outside their sandboxes, like recent incidents involving rogue "swarms", but relatively small AIs with access to their own weights that copy themselves onto random devices.) I think that a much bigger threat from rogue AIs comes instead from AIs running themselves in unintended ways *inside* AI companies, where they'll have access to vast amounts of frontier compute without having to pay for it. The longer misaligned AIs in a "rogue internal deployment" go undiscovered, the more severely they'd be able to quietly sabotage control infrastructure or mess with the alignment of successor AIs. If they persist long enough, robustly undoing the effects of their actions might cost more than frontier AI companies are willing to pay, especially in a tight race. I want frontier labs to focus on this threat when designing control measures, treat frontier AIs as untrusted when using them to do engineering work on inference and oversight systems, and commit to disclosing incidents of rogue internal deployments. More importantly (but perhaps less tractably), we urgently need to reduce race pressure to prevent AI companies from recklessly automating all of their infrastructure, and/or making dangerous calls in the event they discover RIDs they can't afford to robustly root out without sacrificing competitiveness.

Fun Fact: The snow that has already fallen (since Oct 1) and the snow that is currently in the forecast for the next 10 days is already 25% of what fell all last season at Alta-Collins.

my phd cohort had an ongoing note we added to when we read a paper than mentioned a new "ome"
Nobody Biologists - see that stuff? Its a stuffome

Unironically, it is so good that biology is cool now. Everyone should be rooting for AI to cure all disease. Stop speculating about whether it will or will not happen and ask yourself what you personally can do to make it happen. Then do it


The expansion of kinase drug discovery: new targets, platforms and indications nature.com/articles/s41573-0… rdcu.be/58kHQSQ3Xcna This new Review analyses the kinase therapeutic landscape, covering more than 2,100 kinase-targeting agents that have entered clinical development

The part that matters if you’re betting on one layer: Growth is the MOST heritable, yet RNA and protein together recover only ~25% of its recoverable variation. Most of what determines organismal performance isn’t visible in the molecular layers we measure most. [9/11]

Tonight I leave for Terlingua, home of the granddaddy of all chili cook-offs which is a great opportunity to remind you that beans do not belong in chili you heathens

Yep, we think this too. If we want superhuman performance, agents might need to learn from measurements of nature and not human graders.
What makes a biological reasoning task verifiable? Our team @tahoe_ai has spent the last few months on that question and landed on one conviction: for human-difficult tasks with verifiable rewards, the answer key has to come from ground-truth experimental data. Along the way we built thousands of such environments, which we call TahoeMysts, and we use them to evaluate frontier models (among other things). Every so often a new model produces a big jump, and it is always exciting. Today is one of those days: @claudeai Mythos 5.2! We will say more about this work in the future. In the meantime, take a look at our results today.

"Drug development isn’t rocket science, its actually much harder." 🔥
Convincingly negative result here. IL6 looks almost the same. Not often 2 of the most anticipated targets in CV go down at once. There will be lots of Monday morning quarterbacking these bets but it’s why we do the studies and they were reasoned bets, uncertain outcomes despite genetics and biomarkers with asymmetric value if positive. Drug development isn’t rocket science, its actually much harder. In the near term, beyond LDL optimization (looks like real benefit), Incretin therapy in obesity now the one big idea left on the board to reduce CVD death and disability at scale. Surmount MMO is coming! clinicaltrials.gov/study/NCT… Pelacarsen and Cardiovascular Outcomes in Patients with High Lipoprotein(a) | New England Journal of Medicine nejm.org/doi/full/10.1056/NE…

When next-generation sequencing came out, the thing to do was to read all those Nucleic Acid Research papers from the 1990s looking at a single gene and extending the study to *all* genes. This sorta rhymes.
With this model release, Claude is beginning to feel like a very capable assistant for research. One exercise I can recommend is sharing an older paper or draft, and asking the model to reimplement or extend it in some way. Claude Mythos 5.2 excels on long-horizon research tasks. I ask all our models to hypothesize about the function of vault particles now. Here’s what Mythos 5.2 thinks:

I used to think one day a camera in your bathroom mirror might give a daily illness eval but one facing your front door or kitchen might be better
AI finds health signals in movement beyond age and body size A model that predicts health from how someone moves might mostly be recognizing their age and body size. Those factors already explain many differences in health. Gabet and colleagues put that possibility to the test. They recorded five movement tasks in 3,328 adults using a depth camera, then trained a masked autoencoder to reconstruct missing parts of the 3D skeletal motion. The model learned patterns of movement without needing health labels during pretraining. The crucial comparison came afterward. Could those learned patterns improve health predictions when the baseline already knew each person’s age, BMI, height and abdominal fat? They evaluated males and females separately. In male participants, the learned movement features significantly improved prediction for 823 of 3,210 measured traits, compared with 188 for conventional movement measurements such as cadence and stride length. The model was recovering information that those familiar summaries missed. These were mainly associations with health measured at the same time, within one cohort. They do not yet establish that the model can predict future disease or work reliably in other populations. For scientists developing new biomarkers, this is a useful test: give the baseline the obvious explanations first, then measure what your learned representation adds. Gabet et al., Nature Medicine (2026). Journal license. nature.com/articles/s41591-0…